← BlogRelease · AZ-Focus · 22 October 2026

Max effort, less waffle.

Meta's Muse Glimmer 30B lets you choose how hard it thinks, from low to xhigh. Today we're releasing Muse-Glimmer-30B-AZ-Focus, open weights with the second-guessing turned down. You get xhigh results for fewer tokens than the original spends at its default setting.

Max effort, less waffle: Muse Glimmer 30B with AZ-Focus on one DGX Spark (Q8_0, llama.cpp). Mean thinking tokens per answer at high (default): original 4,532, AZ-Focus 3,307 (−27%), correct 102 vs 100 of 128; at xhigh: 5,211 vs 3,956 (−24%), 102 correct each. Correct within an 8k budget at xhigh: 85 vs 91 of 128; hardest (GPQA Diamond + AIME) 14 vs 19 of 32. AZ-Focus at xhigh uses 13% fewer tokens than the original at its default. SWE-bench Verified, first 50: 32 vs 29 resolved.
−27%
thinking at the default reasoning strength (high), 100 vs 102 correct of 128
−13%
thinking at xhigh compared with the original at high: 3,956 vs 4,532 tokens per answer, 102 correct each
32 · 29
SWE-bench Verified tasks resolved of the first 50, AZ-Focus vs the original

Biggest savings on the longest answers

AZ-Focus saves most where Muse Glimmer thinks longest. At the default setting its longest answer was 38,921 tokens, against 53,660 for the original, and on the hardest questions it spends far less time circling: 4,675 thinking tokens on average where it doesn't reach the answer, against 8,200 for the original.

That's where Glimmer's thinking goes. It's already quick on most questions, with a median answer of about 1,300 thinking tokens on our 64-question test set (MMLU-Pro, BIG-Bench Hard, GPQA Diamond and AIME 2025, each question answered twice). The longest 10% of answers account for 58% of all its thinking, and much of it is re-checking work already done: "Wait…", "Hmm…", "Actually…". AZ-Focus still reasons step by step and still checks its work. It just loops less.

Turned up to xhigh, still cheaper than the default

AZ-Focus cuts thinking at both of Muse Glimmer's thinking-heavy settings, by 27% at high and 24% at xhigh. At xhigh it uses fewer tokens than the original does at high, for the same 102 correct answers:

Fig. 1: Thinking tokens per item at each effort settingOriginal AZ-Focus
128 answers
Strength highdefault
4,532 · 79.7%
3,307 · 78.1%−27%
Strength xhigh
5,211 · 79.7%
3,956 · 79.7%−24%
↑ Original's lowest setting
Mean thinking tokens per answer · accuracy, 64 questions answered twice at each setting. The dashed line marks the original at its default (high). Q8_0 GGUFs of both models on the same llama.cpp server (one DGX Spark), temperature 1.0, top_p 0.95, 128k output limit; no answer hit the limit. Accuracy differences of an answer or two are within noise.

If you'd been running Muse Glimmer at high to keep answers quick, AZ-Focus at xhigh gives you the model's most thorough setting at a lower thinking cost.

More correct answers within a budget

Shorter answers pay off most when answers have to come back within a time or token budget. Counting an answer only if it is correct and finished within the budget, at xhigh:

Fig. 2: Answers correct within a token budget, reasoning strength xhighOriginal AZ-Focus AZ-Focus ahead
All 128 answers0204060801002k4k8k16k32kno limittoken budget (thinking + answer)correct617385941011026379919899102Hardest: GPQA Diamond + AIME (32 answers)0204060801002k4k8k16k32kno limittoken budget (thinking + answer)correct371420272841019252627
Same runs as Fig. 1, at xhigh. An answer counts if it is correct and its thinking plus answer fits within the budget.

At an 8k budget AZ-Focus gets 91 answers correct to the original's 85; on the hardest questions, 19 to 14. With no limit the two end level, as they should: AZ-Focus reaches the same answers, sooner.

Just as strong as an agent

Muse Glimmer is built for agentic work, so we checked that AZ-Focus keeps the careful checking a coding agent needs. We ran both models on the first 50 tasks of SWE-bench Verified with mini-SWE-agent, at the default reasoning strength:

SWE-bench Verified, first 50 tasksOriginalAZ-Focus
Tasks resolved29 of 5032 of 50
Finished within the 100-step limit37 of 5041 of 50

AZ-Focus resolved 32 tasks to the original's 29, finished more of them within the step limit, and took slightly fewer steps (67.4 per task on average, against 69.4). On 50 tasks the difference is within noise, so we read it as the same agentic coding, in slightly fewer steps.

Get it

Muse-Glimmer-30B-AZ-Focus is open, under Apache 2.0 like the original, and a drop-in replacement: the same files, chat template, reasoning strengths and vision encoder.

What this does and doesn't show

  • Small sets: 128 answers per setting and 50 coding tasks. Differences of a few answers are noise; the thinking reductions of 24–27% are the signal.
  • Measured with thinking on, at high and xhigh, on Q8_0 GGUFs in llama.cpp with its default top-k (Meta recommends 64; both models were run the same way). Lower strengths, vision, other languages and very long contexts haven't been measured separately.
  • AZ-Focus is designed to change how long the model thinks, not what it will or won't do. We haven't re-run Meta's safety evaluations; Meta's usage policy applies.